[dpdk-users] Tx burst and ordered packets

[Alan Arondel <arondel.alan@gmail.com>]

> >* Are you using RSS on the receiver.
*> >* Packets sent in one burst should go out in order (if not it is a Intel
*> >* driver bug)
*
> [Sidenote: its best practice to reply inline when answering questions on mailing lists.]

Sorry, i'm a beginner with mailing list. No i'm not using RSS.

> >* On Mon, Jun 7, 2021 at 3:09 PM Alan Arondel <Alan.Arondel at enea.com <https://mails.dpdk.org/listinfo/users>> wrote:
*> > > >* > Hello everyone,
*> >* >
*> >* > I just wanted to know if a burst of packets using rte_eth_tx_burst are
*> >* > send in the same order of the parameter struct rte_mbuf ** tx_pkts ?
*> >* > I tried the skeleton sample which use this api and can confirm that the
*> >* > received packets are not in the right order. Probably I'm
doing something
*> >* > wrong somewhere.
*
> Stephen is correct in that the rte_mbuf **tx_pkts order is the order packets
> should be sent on the wire. Indeed if they are not, it is a driver bug, and should
> be fixed.

> Before we jump to that conclusion, can I suggest we look at how packet ordering
> is being measured? Keep in mind packets could be dropped along the way, so
> if we count (N) and (N+1) on the next packet, and then validate ordering by ensuring
> the received packets are exactly (N, N+1), then a drop would suggest re-ordering issues...

Let me explain a bit more what i am doing :

I'm currently using two hosts. These two hosts are connected together
without any device between them on 2 ports (10G interfaces). One of
the host is generating traffic using tcpreplay with a pcap of 12
packets (basicaly http) on one interface and i have a tcpdump on the
other interface. The replay is done only once so I expect to send and
receive back these 12 packets. the other host is hosting the
application with dpdk (skeleton sample). I changed the sample in order
to wait for a burst of 6 packets. Once the burst of the 6 packets is
ready, I use the rte_eth_tx_burst. I displayed the burst on terminal
and also check with gdb and no issue seems to be present with the
order of packets.

I receive correctly the 12 packets but the order seems to be random
inside of the burst which is a bit confusing (not always the same
result). However, I receive correctly the first 6 packets and then the
last 6 packets together.

>* > I use the driver net_i40e.
*>* > Dpdk version 20.11.1
*>* > Compilation without any specific parameters (meson
*>* > configure/ninja/ninja install)
*>* > I cannot find any discussion or information about this.
*
I expect that would work correctly.

>* > Thank you for any reply.
*>* >
*>* > Regards,
*>* > Alan.
*
Regards, -Harry